def objective(trial):
iris = load_iris()
X, y = cudf.DataFrame(pd.DataFrame(
iris.data.astype('float32'))), cudf.DataFrame(
pd.DataFrame(iris.target.astype('float32')))
solver = trial.suggest_categorical("solver", ["qn"])
C = trial.suggest_uniform("C", 0.0, 1.0)
if solver == "qn":
penalty = trial.suggest_categorical("penalty", ["l1", "l2"])
else:
# 'penalty' parameter isn't relevant for this solver,
# so we always specify 'l2' as the dummy value.
penalty = "l2"
classifier = LogisticRegression(max_iter=200,
solver=solver,
C=C,
penalty=penalty)
X_train, X_valid, y_train, y_valid = train_test_split(X, y)
classifier.fit(X_train, y_train)
score = classifier.score(X_valid, y_valid)
return score
def fast_logistic_regression(
X,
y,
*,
penalty: Literal['l1', 'l2', 'elasticnet', 'none'] = 'l2',
C: float = 1.0,
solver: Literal['newton-cg', 'lbfgs', 'liblinear', 'sag',
'saga'] = 'lbfgs',
fit_intercept: bool = True,
l1_ratio: Optional[float] = None,
tol: float = 1e-4,
max_iter: int = 1000,
class_weight: Optional[Dict[str, float]] = None,
n_jobs: Optional[int] = None,
random_state: int = 1,
framework: Literal['auto', 'cuml', 'sklearn'] = 'sklearn',
**kwargs,
) -> LogisticRegression:
"""The cuML LogisticRegression is only faster when n_samples > 100000 given 64
feature dimensions"""
kw = _prepare_kw(locals(), 'X', 'y', 'kwargs', 'framework', **kwargs)
### import
is_cuml = False
if framework == 'sklearn':
LoRe = LogisticRegression
else:
try:
from cuml.linear_model import LogisticRegression as LoRe
is_cuml = True
kw.pop('n_jobs')
kw.pop('random_state')
# if solver not in CUML_SOLVER:
kw['solver'] = 'qn'
except ImportError as e:
LoRe = LogisticRegression
### train
model = LoRe(**kw)
model.fit(X, y)
return model
def run_log_reg(scaled_df):
raw_train_arr = []
raw_test_arr = []
# Init metrics
metrics = ['accuracy', 'f1', 'roc_auc_ovr']
# Set c vals and penalty
C_vals = range(-8, 5)
C_vals = [10**val for val in C_vals]
penalty = ['none', 'l1', 'l2']
# Init params
params = {'penalty': penalty, 'C': C_vals}
# Over five trials
for i in range(5):
# Train test split
X_train, X_test, y_train, y_test = train_test_split(
scaled_df.iloc[:, :-1], scaled_df.y, train_size=5000)
# Init clf
clf = LogisticRegression()
# Init gridsearch and run
search_results = GridSearchCV(clf,
params,
scoring=metrics,
refit=False)
search_results.fit(X_train, y_train)
# Get results and organize
results = pd.DataFrame(search_results.cv_results_['params'])
results['mean_accuracy'] = search_results.cv_results_[
'mean_test_accuracy']
results['mean_f1'] = search_results.cv_results_['mean_test_f1']
results['mean_auc'] = search_results.cv_results_[
'mean_test_roc_auc_ovr']
# Get optimal clfs
opt_acc_inf = results.sort_values(by='mean_accuracy',
ascending=False).iloc[0]
opt_f1_inf = results.sort_values(by='mean_f1', ascending=False).iloc[0]
opt_auc_inf = results.sort_values(by='mean_auc',
ascending=False).iloc[0]
# Init optimal clfs
opt_acc_clf = LogisticRegression(C=opt_acc_inf.C,
penalty=opt_acc_inf.penalty,
max_iter=100000)
opt_f1_clf = LogisticRegression(C=opt_f1_inf.C,
penalty=opt_f1_inf.penalty,
max_iter=100000)
opt_auc_clf = LogisticRegression(C=opt_auc_inf.C,
penalty=opt_auc_inf.penalty,
max_iter=100000)
# Fit clfs
opt_acc_clf.fit(X_train, y_train)
opt_f1_clf.fit(X_train, y_train)
opt_auc_clf.fit(X_train, y_train)
# Get train and test metrics
train_score_acc = opt_acc_clf.score(X_train, y_train)
train_score_f1 = f1_score(y_train, opt_f1_clf.predict(X_train))
train_score_auc = roc_auc_score(y_train, opt_auc_clf.predict(X_train))
test_score_acc = opt_acc_clf.score(X_test, y_test)
test_score_f1 = f1_score(y_test, opt_f1_clf.predict(X_test))
test_score_auc = roc_auc_score(y_test, opt_auc_clf.predict(X_test))
raw_train_arr.append(
[train_score_acc, train_score_f1, train_score_auc])
raw_test_arr.append([test_score_acc, test_score_f1, test_score_auc])
# Create dataframe from results
raw_train_arr = np.array(raw_train_arr).reshape(5, 3)
raw_test_arr = np.array(raw_test_arr).reshape(5, 3)
raw_train_df = pd.DataFrame(data=raw_train_arr,
columns=['accuracy', 'f1', 'auc'])
raw_test_df = pd.DataFrame(data=raw_test_arr,
columns=['accuracy', 'f1', 'auc'])
# Return results
return raw_train_df, raw_test_df
class CumlLRFitter(FitterBase):
def __init__(self,
label='label',
metric='error',
opt: LROpt = None,
max_eval=10):
super(CumlLRFitter, self).__init__(label, metric, max_eval)
if opt is not None:
self.opt = opt
else:
self.opt = LROpt()
self.clf = None
def train(self, train_df, eval_df, params=None):
train_df, eval_df = cudf.DataFrame(train_df), cudf.DataFrame(eval_df)
x_train, y_train, x_eval, y_eval = train_df.drop(columns=[self.label]), train_df[self.label], \
eval_df.drop(columns=[self.label]), eval_df[self.label],
if params is None:
use_params = deepcopy(self.opt_params)
else:
use_params = deepcopy(params)
self.clf = LogisticRegression(**use_params)
self.clf.fit(X=x_train, y=y_train)
preds = self.clf.predict(X=x_eval)
output = self.get_loss(y_pred=preds, y=y_eval)
return output
def search(self, train_df, eval_df):
self.opt_params = dict()
def train_impl(params):
self.train(train_df, eval_df, params)
if self.metric == 'auc':
y_pred = self.clf.predict(eval_df.drop(columns=[self.label]))
else:
y_pred = self.clf.predict(
eval_df.drop(columns=[self.label])).astype(int)
return self.get_loss(eval_df[self.label], y_pred)
self.opt_params = fmin(train_impl,
asdict(self.opt),
algo=tpe.suggest,
max_evals=self.max_eval)
def search_k_fold(self, k_fold, data):
self.opt_params = dict()
def train_impl_nfold(params):
loss = list()
for train_id, eval_id in k_fold.split(data):
train_df = data.iloc[train_id, :]
eval_df = data.iloc[eval_id, :]
self.train(train_df, eval_df, params)
if self.metric == 'auc':
y_pred = self.clf.predict(
eval_df.drop(columns=[self.label]))
else:
y_pred = self.clf.predict(
eval_df.drop(columns=[self.label])).astype(int)
loss.append(self.get_loss(eval_df[self.label], y_pred))
return np.mean(loss)
self.opt_params = fmin(train_impl_nfold,
asdict(self.opt),
algo=tpe.suggest,
max_evals=self.max_eval)
def train_k_fold(self,
k_fold,
train_data,
test_data,
params=None,
drop_test_y=True):
acc_result = list()
train_pred = cudf.Series(np.empty(train_data.shape[0]))
test_pred = cudf.Series(np.empty(test_data.shape[0]))
if drop_test_y:
dtest = test_data.drop(columns=self.label)
else:
dtest = test_data
for train_id, eval_id in k_fold.split(train_data):
train_df = train_data.iloc[train_id, :]
eval_df = train_data.iloc[eval_id, :]
self.train(train_df, eval_df, params)
train_pred[eval_id] = self.clf.predict_proba(
eval_df.drop(columns=self.label)).iloc[:, 1].values
if self.metric == 'auc':
y_pred = self.clf.predict(eval_df.drop(columns=[self.label]))
else:
y_pred = self.clf.predict(
eval_df.drop(columns=[self.label])).astype(int)
acc_result.append(self.get_loss(eval_df[self.label], y_pred))
test_pred += self.clf.predict_proba(dtest).iloc[:, 1]
test_pred /= k_fold.n_splits
return train_pred, test_pred, acc_result
def rank_genes_groups(
X,
labels, # louvain results
var_names,
groups=None,
reference='rest',
n_genes=100,
**kwds,
):
"""
Rank genes for characterizing groups.
Parameters
----------
X : cupy.ndarray of shape (n_cells, n_genes)
The cellxgene matrix to rank genes
labels : cudf.Series of size (n_cells,)
Observations groupings to consider
var_names : cudf.Series of size (n_genes,)
Names of genes in X
groups : Iterable[str] (default: 'all')
Subset of groups, e.g. ['g1', 'g2', 'g3'], to which comparison
shall be restricted, or 'all' (default), for all groups.
reference : str (default: 'rest')
If 'rest', compare each group to the union of the rest of the group.
If a group identifier, compare with respect to this group.
n_genes : int (default: 100)
The number of genes that appear in the returned tables.
"""
#### Wherever we see "adata.obs[groupby], we should just replace w/ the groups"
import time
start = time.time()
# for clarity, rename variable
if groups == 'all':
groups_order = 'all'
elif isinstance(groups, (str, int)):
raise ValueError('Specify a sequence of groups')
else:
groups_order = list(groups)
if isinstance(groups_order[0], int):
groups_order = [str(n) for n in groups_order]
if reference != 'rest' and reference not in set(groups_order):
groups_order += [reference]
if (reference != 'rest' and reference not in set(labels.cat.categories)):
cats = labels.cat.categories.tolist()
raise ValueError(
f'reference = {reference} needs to be one of groupby = {cats}.')
groups_order, groups_masks = select_groups(labels, groups_order)
original_reference = reference
n_vars = len(var_names)
# for clarity, rename variable
n_genes_user = n_genes
# make sure indices are not OoB in case there are less genes than n_genes
if n_genes_user > X.shape[1]:
n_genes_user = X.shape[1]
# in the following, n_genes is simply another name for the total number of genes
n_genes = X.shape[1]
n_groups = groups_masks.shape[0]
ns = cp.zeros(n_groups, dtype=int)
for imask, mask in enumerate(groups_masks):
ns[imask] = cp.where(mask)[0].size
if reference != 'rest':
ireference = cp.where(groups_order == reference)[0][0]
reference_indices = cp.arange(n_vars, dtype=int)
rankings_gene_scores = []
rankings_gene_names = []
# Perform LogReg
# if reference is not set, then the groups listed will be compared to the rest
# if reference is set, then the groups listed will be compared only to the other groups listed
from cuml.linear_model import LogisticRegression
reference = groups_order[0]
if len(groups) == 1:
raise Exception(
'Cannot perform logistic regression on a single cluster.')
grouping_mask = labels.astype('int').isin(cudf.Series(groups_order))
grouping = labels.loc[grouping_mask]
X = X[grouping_mask.
values, :] # Indexing with a series causes issues, possibly segfault
y = labels.loc[grouping]
clf = LogisticRegression(**kwds)
clf.fit(X.get(), grouping.to_array().astype('float32'))
scores_all = cp.array(clf.coef_).T
for igroup, group in enumerate(groups_order):
if len(groups_order) <= 2: # binary logistic regression
scores = scores_all[0]
else:
scores = scores_all[igroup]
partition = cp.argpartition(scores, -n_genes_user)[-n_genes_user:]
partial_indices = cp.argsort(scores[partition])[::-1]
global_indices = reference_indices[partition][partial_indices]
rankings_gene_scores.append(scores[global_indices].get(
)) ## Shouldn't need to take this off device
rankings_gene_names.append(var_names[global_indices].to_pandas())
if len(groups_order) <= 2:
break
groups_order_save = [str(g) for g in groups_order]
if (len(groups) == 2):
groups_order_save = [g for g in groups_order if g != reference]
print("Ranking took (GPU): " + str(time.time() - start))
start = time.time()
scores = np.rec.fromarrays(
[n for n in rankings_gene_scores],
dtype=[(rn, 'float32') for rn in groups_order_save],
)
names = np.rec.fromarrays(
[n for n in rankings_gene_names],
dtype=[(rn, 'U50') for rn in groups_order_save],
)
print("Preparing output np.rec.fromarrays took (CPU): " +
str(time.time() - start))
print("Note: This operation will be accelerated in a future version")
return scores, names, original_reference
def rank_genes_groups(
X,
labels, # louvain results
var_names,
groupby=str,
groups=None,
reference='rest',
n_genes=100,
key_added=None,
layer=None,
**kwds,
):
#### Wherever we see "adata.obs[groupby], we should just replace w/ the groups"
import time
start = time.time()
# for clarity, rename variable
if groups == 'all':
groups_order = 'all'
elif isinstance(groups, (str, int)):
raise ValueError('Specify a sequence of groups')
else:
groups_order = list(groups)
if isinstance(groups_order[0], int):
groups_order = [str(n) for n in groups_order]
if reference != 'rest' and reference not in set(groups_order):
groups_order += [reference]
if (reference != 'rest' and reference not in set(labels.cat.categories)):
cats = labels.cat.categories.tolist()
raise ValueError(
f'reference = {reference} needs to be one of groupby = {cats}.')
groups_order, groups_masks = select_groups(labels, groups_order)
original_reference = reference
n_vars = len(var_names)
# for clarity, rename variable
n_genes_user = n_genes
# make sure indices are not OoB in case there are less genes than n_genes
if n_genes_user > X.shape[1]:
n_genes_user = X.shape[1]
# in the following, n_genes is simply another name for the total number of genes
n_genes = X.shape[1]
n_groups = groups_masks.shape[0]
ns = cp.zeros(n_groups, dtype=int)
for imask, mask in enumerate(groups_masks):
ns[imask] = cp.where(mask)[0].size
if reference != 'rest':
ireference = cp.where(groups_order == reference)[0][0]
reference_indices = cp.arange(n_vars, dtype=int)
rankings_gene_scores = []
rankings_gene_names = []
rankings_gene_logfoldchanges = []
rankings_gene_pvals = []
rankings_gene_pvals_adj = []
# if 'log1p' in adata.uns_keys() and adata.uns['log1p']['base'] is not None:
# expm1_func = lambda x: np.expm1(x * np.log(adata.uns['log1p']['base']))
# else:
# expm1_func = np.expm1
# Perform LogReg
# if reference is not set, then the groups listed will be compared to the rest
# if reference is set, then the groups listed will be compared only to the other groups listed
from cuml.linear_model import LogisticRegression
reference = groups_order[0]
if len(groups) == 1:
raise Exception(
'Cannot perform logistic regression on a single cluster.')
grouping_mask = labels.astype('int').isin(cudf.Series(groups_order))
grouping = labels.loc[grouping_mask]
X = X[grouping_mask.
values, :] # Indexing with a series causes issues, possibly segfault
y = labels.loc[grouping]
clf = LogisticRegression(**kwds)
clf.fit(X.get(), grouping.to_array().astype('float32'))
scores_all = cp.array(clf.coef_).T
for igroup, group in enumerate(groups_order):
if len(groups_order) <= 2: # binary logistic regression
scores = scores_all[0]
else:
scores = scores_all[igroup]
partition = cp.argpartition(scores, -n_genes_user)[-n_genes_user:]
partial_indices = cp.argsort(scores[partition])[::-1]
global_indices = reference_indices[partition][partial_indices]
rankings_gene_scores.append(scores[global_indices].get(
)) ## Shouldn't need to take this off device
rankings_gene_names.append(var_names[global_indices].to_pandas())
if len(groups_order) <= 2:
break
groups_order_save = [str(g) for g in groups_order]
if (len(groups) == 2):
groups_order_save = [g for g in groups_order if g != reference]
print("Ranking took (GPU): " + str(time.time() - start))
start = time.time()
scores = np.rec.fromarrays(
[n for n in rankings_gene_scores],
dtype=[(rn, 'float32') for rn in groups_order_save],
)
names = np.rec.fromarrays(
[n for n in rankings_gene_names],
dtype=[(rn, 'U50') for rn in groups_order_save],
)
print("Preparing output np.rec.fromarrays took (CPU): " +
str(time.time() - start))
print("Note: This operation will be accelerated in a future version")
return scores, names, original_reference